On Jan 15, 2008, at 5:54 PM, Brendan Gregg wrote:

> G'Day Folks,
>
> I'm testing out the prototype NFSv3 DTrace provider, which is  
> working well so far (and for anyone who missed the news, the NFSv4  
> provider was putback to Solaris Nevada last year and works great).
>
> One of my scripts, nfsv3snoop.d, prints filename, offset and size.   
> While reading through a file remotely (using cksum), it traced the  
> following:
>
> # ./nfsv3iosnoop2.d
> TIME(us)         CLIENT             OP OFFSET    BYTES PATHNAME
> 738601899        192.168.17.75      Rs 0          4096 /export/ 
> stuff/words
> 738613529        192.168.17.75      Rd 0          4096 /export/ 
> stuff/words
> 738614258        192.168.17.75      Rs 4         28672 /export/ 
> stuff/words
> 738627681        192.168.17.75      Rd 4         28672 /export/ 
> stuff/words
> 738628599        192.168.17.75      Rs 32        32768 /export/ 
> stuff/words
> 738628683        192.168.17.75      Rd 32        32768 /export/ 
> stuff/words
> 738629509        192.168.17.75      Rs 96        32768 /export/ 
> stuff/words
> 738629597        192.168.17.75      Rd 96        32768 /export/ 
> stuff/words
> 738629800        192.168.17.75      Rs 128       32768 /export/ 
> stuff/words
> 738629886        192.168.17.75      Rd 128       32768 /export/ 
> stuff/words
> 738630316        192.168.17.75      Rs 160       32768 /export/ 
> stuff/words
> 738630439        192.168.17.75      Rd 160       32768 /export/ 
> stuff/words
> 738630748        192.168.17.75      Rs 64        32768 /export/ 
> stuff/words
> 738630945        192.168.17.75      Rd 64        32768 /export/ 
> stuff/words
> 738632525        192.168.17.75      Rs 192       12288 /export/ 
> stuff/words
> 738632577        192.168.17.75      Rd 192       10055 /export/ 
> stuff/words
> ^C
>
> Rs == Read start, Rd == Read done.  OFFSET is in Kbytes, BYTES is  
> in bytes.
>
> Note how the read requests skip the 64 Kbyte offset, and then  
> return to it 1.2 ms later?  Is there a reason why NFSv3 would do  
> this? ... I can use DTrace to investigate - someone might already  
> know what's up.

Scheduling of the various NFS threads handling the reads?

There is nothing in the NFS server that would try to do this
so it is a result of other unintentional interactions. :-)

Spencer

>
> This is on a single CPU server, and is repeatable.
>
> Brendan
>
>
> This message posted from opensolaris.org
> _______________________________________________
> nfs-discuss mailing list
> nfs-discuss at opensolaris.org


Reply via email to